Memory is what gives agents continuity, personalization, and depth—but managing it is far from simple. This talk explores current best practices in agentic memory, from short-term working memory with summarization to long-term stores with semantic retrieval and metadata filters. We’ll cover how modern systems handle promotion, deduplication, and background consolidation, as well as how protocols like MCP expose memory as first-class tools. We'll illustrate design trade-offs and practical strategies for building memory-aware agents in Java.

Jfokus

A conference talk by Brian Sam-Bodden from Redis exploring how to build an agentic memory management server. Drawing from neuroscience, psychology, and computer science, the talk covers how human memory principles (encoding, storage, retrieval, forgetting) can be applied to AI systems. Six memory types are introduced: semantic, episodic, procedural, entity, and others. The Redis Spring AI Memory Server is demonstrated, showing how memories persist across sessions using hybrid search (BM25 + vector search with reciprocal rank fusion). Key concepts include long-term vs short-term memory, memory eviction policies, conflict resolution, background consolidation processes, and exposing the system as an MCP server for use with Claude or ChatGPT.

Secrets of Agentic Memory Management, by Brian Sam-Bodden